Current knowledge about impact of colour models on performance of convolutional neural network is investigated in the first part of this thesis. The experiment based on obtained knowledge is conducted in the second part. Six colour models HSV, CIE 1931 XYZ, CIE 1976 L*a*b*, YIQ a YCbCr and deep convolutional neural network ResNet-101 are used. RGB colour model achieved the highest classification accuracy, whereas HSV color model has the lowest accuracy in this experiment.

This paper presents basic principles of optical coherence tomography, review of applications and basic categorization of these systems. Paper also deals with the typical properties of images from optical coherence tomography, especially speckle pattern. This paper also provides an overview of the origin of speckle noise and utilization of its dependence on microstructure of probed tissue for image classification based on textural analysis. Experimental part of this paper consists of phantom preparation, data acquisition by OCT system, implementation of speckle analysis in MATLAB and of testing of its functionality on standard textural dataset and also on acquired image phantom data. Speckle analysis is used for phantom image data segmentation.

This thesis deals with image object detection and its classification into classes. Classification is provided by models of framework for deep learning BVLC/Caffe. Object detection is provided by AlpacaDB/selectivesearch and belltailjp/selective_search_py algorithms. One of results of this thesis is modification and usage of deep convolutional neural network AlexNet in BVLC/Caffe framework. This model was trained with precision 51,75% for classification into 1 000 classes. Then it was modified and trained for classification into 20 classes with precision 75.50%. Contribution of this thesis is implementation of graphical interface for object detction and their classification into classes, which is implemented as aplication based on web server in Python language. Aplication integrates object detection algorithms mentioned abowe with classification with help of BVLC/Caffe. Resulting aplication can be used for both object detection (and classification) and for fast verification of any classification model of BVLC/Caffe. This aplication was published on server GitHub under license Apache 2.0 so it can be further implemented and used.

In the past few years, complex neural networks have achieved state of the art results in image classification. However, training these models requires large amounts of labelled data. Whereas unlabelled images are often readily available in large quantities, obtaining l abels takes considerable human effort. Active learning reduces the required labelling effort by selecting the most informative instances to label. The most popular active learning query strategy framework, uncertainty sampling, uses uncertainty estimates of the model being trained to select instances for labelling. However, modern classification neural networks often do not provide good uncertainty estimates. Baye sian neural networks model uncertainties over model parameters, which can be used to obtain uncertainties over model predictions. Exact Bayesian inference is intractable for neural networks, however several approximate methods have been proposed. We experiment with three such methods using various uncertainty sampling active learning query strategies.

Sundarban Reserved Forest in Bangladesh is playing a significant role in local and national economy and is providing protection to the coastline as well as to the indigenous people. During the past decades and also in recent time this forest was heavily disturbed by human intervention in many aspects. As a consequence the resources of the forest are fragmenting, shrinking and declining, which in turn leads to an increasing failure of satisfying increasing demands both at local and national levels. Therefore accurate and continuously updated spatial information is needed for optimising forest management and environmental planning on both levels to support the fulfilment of urgent needs of sustainability of the forest. Considering the specific topography and the poor accessibility of the forest versus the task of collecting information, remote sensing is an attractive, if not the only means of obtaining sound full-coverage spatial information on forest cover of Sundarban. This research used medium resolution Landsat ETM data of November 2000 and Landsat TM data of January 1989 to assess and monitor the forest for 1. Identification of the operational tools for mapping and monitoring the forest as well as on the examination of the reliability of the application of multitemporal satellite remote sensing data for building spatial databases on forest cover in Sundarban. 2. Based on the existing management plan of the forest as well as the spectral properties of Landsat ETM imagery a level III classification system was developed. 3. This classification strategy was tested by applying several methods to achieve the classification result with the highest accuracy and thus to build the most reliable methodology for mapping forest cover in Sundarban. 4. Forest cover change was assessed for the period of eleven years. Significant changes have been observed due to illegal removal of trees from the forest although a governmental moratorium on banning timber extraction exists since 1989. 5. Development of an operational monitoring scheme by means of multitemporal satellite imagery analysis, which will allow concerned authorities to set up sustainable and appropriate monitoring of the Sundarban Reserved Forest. / Das Schutzgebiet des Sundarban Mangrovenwaldes in Bangladesh spielt eine entscheidende Rolle in Hinsicht auf nationale und lokale sozio-ökonomische und sozio-ökologische Aspekte. Das Waldgebiet stabilisiert nicht nur die Küstenlinie, sondern schützt auch die Bevölkerung vor den Einflüssen von Flutkatastrophen. Durch menschlichen Einfluss wurde die Region während der letzten Jahrzehnte mehr und mehr unmittelbar gestört. Der Rückgang des Ertrags an Ressourcen aus dem Wald führte zu wachsender Unzufriedenheit in der von diesen Nutzungs-möglichkeiten abhängigen Bevölkerung. Um eine Optimierung des Waldmanagements durchführen zu können, werden kontinuierliche und genaue raumbezogene Daten benötigt. Betrachtet man die spezifische Topographie und die schlechte Zugänglichkeit der Waldgebiete, so bietet die Fernerkundung eine attraktive Möglichkeit, raumbezogene Informationen für die großen Flächen des Sundurban Mangrovenwaldes zu erfassen. Zur Analyse und Überwachung der Waldgebiete wurden zwei Satellitenbild-Datensätze mit mittlerer Auflösung verwendet, und zwar Landsat ETM Daten aus dem Jahre 2000 (November) sowie Landsat TM Daten aus dem Jahre 1989 (Januar). Die zentralen Aktivitäten im Rahmen der Bearbeitung der Dissertation beziehen sich auf 1. die Identifikation der notwendigen Werkzeuge für eine erfolgreiche Kartierung und Überwachung der Waldgebiete sowie Untersuchung der Zuverlässigkeit multi-temporaler Fernerkundungsdaten für den Aufbau einer Datenbasis für die Kartierung von Waldbedeckungsarten im Untersuchungsgebiet des Sunderban Mangroven-waldes, 2. die Entwicklung eines Klassifikationssystems nach dem USGS-Schlüssel (Auflösungsebene III) auf Grundlage des existierenden Managementplanes und der spektralen Qualität der Landsat ETM Satellitenbilddaten, 3. den Test der Klassifikationsstrategie durch Adaption unterschiedlicher Methoden und Optimierung in bezug auf Erzielung eines Ergebnisses in maximal erreichbarer Genauigkeit als Ausgangspunkt für den Aufbau einer Methodologie zum Monitoring des Sunderban Mangrovenwaldes, 4. die Extraktion der Veränderungen der Waldbedeckung über ein Zeitintervall von 11 Jahren mit weitreichenden Erkenntnissen zur Dynamik der Degradations-effekte, die hauptsächlich durch illegales Fällen trotz Verbot durch ein Regierungs-memorandum seit 1989 beschleunigt wird, 5. die Entwicklung einer operationellen Monitoring-Struktur mit Hilfe von multi-temporaler Satellitenbildanalyse für ein nachhaltiges und angepasstes raumbezo-genes Management des Sunderban-Mangrovenwaldes.

Rapporten beskriver utvecklingsarbetet och resultatet från utvecklingen av en robotplattform vid namn AIDA (AI Design Assistant), som utvecklades åt Institutionen för datavetenskap vid Linköpings universitet. Plattformen består av en robotarm som utgörs av sex stycken servomotorer, som är anslutna till en enkortsdator. En Android-surfplatta sitter integrerad på robotarmen och har en applikation installerad som utgör användargränssnittet. Tre huvudsakliga funktioner för plattformen utvecklades. Dessa funktioner är objektigenkänning, objektspårning och taligenkänning. Objektigenkänningen kan klassificera fyra olika fruktsorter, objektspårningen kan spåra objekt och följa dem med robotarmen genom inverskinematik, och taligenkänningen kan transkribera tal till text och svara på kommandon. Utifrån resultatet och diskussionen härleds slutsatser över fyra frågeställningar relaterade till utvecklingsarbetet. Projektet utfördes som en del av kursen TDDD96 Kandidatprojekt i programvaruutveckling, och varje projektmedlem har även skrivit ett individuellt bidrag till rapporten som behandlar områden kopplade till projektarbetet. / This report describes the development process and the resulting product from the development of a robot platform named AIDA (AI Design Assistant), that was developed on a request from the Department of Computer and Information Science at Linköping University. The platform consists of a robot arm that is made up by six servo motors connected to a single-board computer. An Android tablet is attached to the robot arm and has an application installed which constitutes the user interface. Three main functions were developed for the platform. These functions constitute object recognition, object tracking, and speech recognition. The object recognition module can classify four different types of fruit, the object tracking module can track objects and follow them by moving the robot arm using inverse kinematics, and the speech recognition module can transcribe speech to text and respond to  audible commands. Conclusions over four questions related to the development of the product are derived from the results and discussion chapters of the report. The project was conducted as a part of the course TDDD96 Software Engineering – Bachelor Project, and each project member has produced an individual contribution to the report which covers subjects related to the project.

Chromosome classification is one of the essential tasks in karyotyping to diagnose genetic abnormalities like some types of cancers and Down syndrome. Deep convolutional neural networks have been widely used in this task, and the accuracy of classification models is exceptionally critical to such sensitive medical diagnoses. However, it is not always possible to meet the expected accuracy rates for diagnosis. So, it is vital to tell how certain or uncertain a model is with its decision. In our work, we use two metrics, entropy and variance, as uncertainty measurements. Moreover, three additional metrics, fail rate, workload, and tolerance range, are used to measure uncertainty metrics’ quality. Four different non-Bayesian methods: deep ensembles, snapshot ensembles, Test Time Augmentation, and Test Time Dropout, are used in experiments. A negative correlation is observed between the accuracy and the uncertainty estimation; the higher the accuracy of the model, the lower the uncertainty. Densenet121 with deep ensembles as the uncertainty evaluation method and variance as the uncertainty metric gives the best outcomes. Densenet121 provides a wider tolerance range and better separation between uncertain and certain predictions. / Kromosomklassificering är en av de viktigaste uppgifterna i Karyotyping för att diagnostisera genetiska abnormiteter som vissa typer av cancer och Downs syndrom. Deep Convolutional Neural Networks har använts i stor utsträckning i denna uppgift, och noggrannheten hos klassificeringsmodeller är exceptionellt kritisk för sådana känsliga medicinska diagnoser. Det är dock inte alltid möjligt att uppfylla de förväntade noggrannhetsgraderna för diagnos. Så det är viktigt att berätta hur säker eller osäker en modell är med sitt beslut. Forskning har gjorts för att uppskatta osäkerheten med bayesiska metoder och icke-bayesiska neurala nätverk, medan lite är känt om kvaliteten på osäkerhetsuppskattningar. I vårt arbete använder vi två mått, entropi och varians, som osäkerhetsmätningar. Dessutom används ytterligare tre mätvärden, felfrekvens, arbetsbelastning och toleransintervall för att mäta osäkerhetsmätarnas kvalitet. Fyra olika icke-bayesiska metoder: djupa ensembler, ögonblicksbilder, Test Time Augmentation (TTA) och Test Time Dropout (TTD) används i experiment. En negativ korrelation observeras mellan noggrannheten och osäkerhetsuppskattningen; ju högre noggrannhet modellen är, desto lägre är osäkerheten. Densenet121 med djupa ensembler som osäkerhetsutvärderingsmetod och varians som osäkerhetsmätvärdet ger de bästa resultaten. De ger ett bredare toleransintervall och bättre separation mellan osäkra och vissa förutsägelser.

The population of northern pike in the Baltic sea has seen a drasticdecrease in numbers in the last couple of decades. The reasons for this are believed to be many, but the majority of them are most likely anthropogenic. Today, many measures are being taken to prevent further decline of pike populations, ranging from nutrient runoff control to habitat restoration. This inevitably gives rise to the problem addressed in this project, namely: how can we best monitor pike populations so that it is possible to accurately assess and verify the effects of these measures over the coming decades? Pike is currently monitored in Sweden by employing expensive and ineffective manual methods of individual marking of pike by a handful of experts. This project provides evidence that such methods could be replaced by a Convolutional Neural Network (CNN), an automatic artificial intelligence system, which can be taught how to identify pike individuals based on their unique patterns. A neural net simulates the functions of neurons in the human brain, which allows it to perform a range of tasks, while a CNN is a neural net specialized for this type of visual recognition task. The results show that the CNN trained in this project can identify pike individuals in the provided data set with upwards of 90% accuracy, with much potential for improvement.

Indiana University-Purdue University Indianapolis (IUPUI) / Research on image classification has grown rapidly in the field of machine learning. Many methods have already been implemented for image classification. Among all these methods, best results have been reported by neural network-based techniques. One of the most important steps in automated image classification is feature extraction. Feature extraction includes two parts: feature construction and feature selection. Many methods for feature extraction exist, but the best ones are related to deep-learning approaches such as network-in-network or deep convolutional network algorithms. Deep learning tries to focus on the level of abstraction and find higher levels of abstraction from the previous level by having multiple layers of hidden layers. The two main problems with using deep-learning approaches are the speed and the number of parameters that should be configured. Small changes or poor selection of parameters can alter the results completely or even make them worse. Tuning these parameters is usually impossible for normal users who do not have super computers because one should run the algorithm and try to tune the parameters according to the results obtained. Thus, this process can be very time consuming. This thesis attempts to address the speed and configuration issues found with traditional deep-network approaches. Some of the traditional methods of unsupervised learning are used to build an automated image-classification approach that takes less time both to configure and to run.

No description available.